Nonword repetition depends on the frequency of sublexical representations at different grain sizes: Evidence from a multi-factorial analysis
Introduction
The accuracy with which people repeat non-existing, but phonologically possible words, such as kipser, is a remarkably good index of their language-related abilities. Performance in this very simple task, known as nonword repetition test (henceforth NWR), has proved an important predictor of novel word learning in both native and foreign language acquisition, as indexed by correlations of NWR scores with vocabulary size (Baddeley et al., 1998, Farnia and Geva, 2011, Gathercole and Baddeley, 1989, Gathercole et al., 1997, Gathercole and Masoura, 2003, Gathercole et al., 1992, Jarrold et al., 2009, Service, E., & Kohonen, V, 1995). Clearly, the NWR task emulates the situation in which learners encounter new lexical items for the first time. In both cases, the learner is presented with new verbal material that needs to be perceived, processed and then repeated. Hence, finding the best predictors of NWR might help to identify key factors entailed in word-learning.
NWR task has also proved effective in differentiating children diagnosed with language disorders such as Specific Language Impairment (SLI) from typically developing children (see Estes, Evans, & Else-Quest, 2007 for a metanalysis). SLI is an umbrella term for a group of language impairments that do not result from any general cognitive deficit (Joanisse and Seidenberg, 1998, Joanisse and Seidenberg, 2003, Leonard, 2014). It is characterized by a range of symptoms including problems with grammar (processing of syntactically complex sentences and applying morphological rules; Bortolini, Caselli, & Leonard, 1997), word learning (Morley et al., 1955, Trauner et al., 2000) and phonology (Elliott, Hammer, & Scholl, 1989). NWR has been considered in many studies as a potential diagnostic tool for SLI (see e.g. Estes et al., 2007, Conti-Ramsden and Botting, 2001, Weismer et al., 2000) and it seems that NWR performance is directly related to the core deficit underlying SLI. Therefore, by illuminating mechanisms underlying performance in the NWR task, we should be able to tap into the nature of SLI itself.
The NWR task was initially used as a test of phonological short-term memory (henceforth pSTM) in studies of vocabulary development (Gathercole and Baddeley, 1989, Gathercole and Baddeley, 1990a, Gathercole and Baddeley, 1990b). The authors of these studies argued that NWR was a purer measure of pSTM than, for instance, the traditionally used digit span, because it does not involve processing of any lexical information (Gathercole & Baddeley, 1990b). However, very early on the claim that NWR is a test of pSTM was brought into question. It was pointed out that the task might also involve phonological and articulatory skills, as well as lexical knowledge (Bowey, 1996, Bowey, 2001, Snowling et al., 1991). Even the authors of the early NWR papers conceded that the measure is a complex one and that it measures skills and knowledge beyond pSTM (Gathercole, 1995, Gathercole, 2006). This has led to extensive debate about what NWR actually measures, and extensive research investigating the factors that influence NWR performance as evidence of the skills and knowledge involved.
However, the results of most previous studies have been limited by their underlying methodological approach, with few, and often just one, factor being manipulated in a factorial design. These factors are then attributed to a specific type of cognitive representation or process. Some of these factors correspond to mechanisms (phonological STM, articulatory dexterity, perceptual acuity) and some to representations (lexical and sublexical knowledge) hypothesized to underlie NWR. Each will be briefly reviewed below.
It has been argued that participants who can hold more information in their pSTM can maintain a temporary representation of the nonword long enough to repeat it, resulting in better NWR performance (Gathercole and Baddeley, 1989, Gathercole and Baddeley, 1990a). Participants with poorer pSTM are not able to maintain the nonword in memory and thus have problems with repeating it. The idea that NWR provides an index of pSTM capacity is supported by two classes of findings. First, participants repeat short nonwords more accurately than long nonwords which tax pSTM to a greater degree (e.g. Gathercole and Baddeley, 1989, Stokes et al., 2006, Weismer et al., 2000). Second, NWR correlates with tasks traditionally associated with phonological memory, namely digit span and the immediate serial recall tasks (Archibald and Gathercole, 2006, Gathercole and Baddeley, 1990b).
Length is by no means the only factor to affect NWR performance, indicating that NWR involves more than a purely quantitative capacity. Many researchers have pointed out that long-term memory may contribute to NWR and participants’ knowledge of the language might be a key factor determining accuracy of NWR (Bowey, 1996, Bowey, 2001, Jones, 2011, Jones and Witherstone, 2011, Jones et al., 2010, Metsala, 1999, Snowling et al., 1991). In this case, learners who are more proficient in a given language should be better at repeating nonwords which resemble real words in this language. This idea is supported by findings showing that NWR accuracy correlates with vocabulary size in L1 (Bowey, 1996, Bowey, 2001, Gathercole et al., 1992, Gathercole and Baddeley, 1989) and in L2 (Gathercole and Masoura, 2003, Masoura and Gathercole, 1999, Service, E., & Kohonen, V, 1995). Moreover, participants are more accurate in repeating highly wordlike nonwords (Archibald and Gathercole, 2006, Gathercole, 1995, Munson et al., 2005). However, these findings may reflect different kinds of linguistic knowledge. Zooming in on more specific parameters of nonwords might help to pinpoint the most relevant aspects of this knowledge.
Some authors have proposed that repeating nonwords relies on lexical knowledge and that participants use phonological representations of whole word forms as an aid in temporarily representing the nonword in pSTM (Gathercole, 2006, Roodenrys and Hinton, 2002). They argue that the presentation of a nonword partially activates representations of its lexical neighbours – phonologically similar words that are already known to the learner. This enables more efficient representation of the nonword in pSTM. Phonological neighbours can also be used to repair a decaying memory trace of a particular nonword in pSTM (a process known as lexical redintegration; Brown and Hulme, 1995, Hulme et al., 1991, Hulme et al., 1997, Hulme et al., 1999, Roodenrys and Hinton, 2002, Roodenrys et al., 2002, Schweickert, 1993). The notion that lexical representations support NWR is in line with studies showing that nonwords coming from dense phonological neighbourhoods are repeated faster and more accurately than those coming from sparse phonological neighbourhoods (Janse and Newman, 2013, Roodenrys and Hinton, 2002, Vitevitch and Luce, 1998, Vitevitch and Luce, 1999, Vitevitch and Luce, 2005).1 It is also supported by studies showing that nonwords containing real words (e.g. bathesis) lead to higher repetition accuracy than nonwords that do not contain embedded words (e.g. fathesis; Dollaghan et al., 1993, Dollaghan et al., 1995).
Nonword repetition may be supported not only by representations at the lexical level but also by representations at a sublexical level. Sublexical representations contain phoneme combinations occurring more or less frequently in a given language. The existence of this level of phonological knowledge has been proposed by frameworks based on the Adaptive Resonance Theory (Grossberg, 1986, Gathercole et al., 1997, Vitevitch and Luce, 1998, Vitevitch and Luce, 1999), the EPAM-VOC and CLASSIC theories (Jones et al., 2007, Jones et al., 2014, Jones, 2016), and by the Lexical Restructuring family of theories (Bowey, 2001, Metsala, 1999). The above-mentioned theories (with the exception of Lexical Restructuring) propose that the presentation of a nonword activates representations of “chunks” (or sequences) of phonemes in long-term memory that are present in this nonword. This activation, in turn, facilitates nonword repetition in similar ways to those that have been proposed for lexical representation: more efficient representation of nonwords in phonological STM and/or sublexical redintegration. The facilitative role of sublexical representations in NWR performance is supported by studies showing that nonwords with high phonotactic probability (i.e. containing combinations of phonemes that are typical for the language) are repeated faster (Vitevitch and Luce, 1998, Vitevitch and Luce, 1999, Vitevitch and Luce, 2005) or more accurately (Coady and Aslin, 2004, Edwards et al., 2004, Gathercole et al., 1999, Majerus et al., 2004, Messer et al., 2010, Munson et al., 2005, Munson et al., 2005, Roodenrys and Hinton, 2002, Thorn et al., 2005, Zamuner, 2009, Zamuner et al., 2004) than nonwords with low phonotactic probability. In all of these studies, phonotactic probability was indexed using phonemic bigram frequency, i.e. the mean frequency of all pairs of adjacent phonemes (or letters) occurring in the nonword.
However, phonemic bigram frequency is not the only possible measure of phonotactic probability. Sublexical representations do not necessarily consist solely of two-phoneme combinations. In fact, EPAM-VOC and CLASSIC theories (Jones et al., 2007, Jones et al., 2014, Jones, 2016) assume that language learners store sublexical representations of varied lengths and that the greater the experience with a particular language, the longer the sequences of phonemes stored. This raises the possibility that a measure of phonotactic probability taking into consideration the frequency of phoneme sequences at many different lengths (phonemic bigrams, trigrams, and so on) might be a better predictor of NWR than simple phonemic bigram frequency. Another possibility is that the most rudimentary units of sublexical representation are syllables or subsyllabic elements such as onsets or codas rather than phonemes. Syllables and subsyllabic elements are typically proposed alongside phonemes as units of speech perception and production in phonology (Cote, 2012, Zec, 2007) and some studies suggest that these units might be indeed more basic and natural than syllables (Anthony et al., 2003, Morais et al., 1979, Massaro, 1987, Pierrehumbert and Nair, 1995, Read et al., 1986, Treiman, 1983, Ziegler and Goswami, 2005)
Apart from pSTM and language-specific knowledge, NWR performance may also depend on articulatory difficulty and hence oromotor dexterity, as indicated by two types of evidence. First, NWR performance is correlated with scores on tasks testing oromotor skills in which participants are asked to repeat complex oral movement sequences after a model (Krishnan et al., 2013, Stark and Blackwell, 1997). Second, nonwords containing consonant clusters, structures that are deemed to be phonologically complex and difficult to articulate, are less likely to be repeated correctly in typically developing children (Archibald and Gathercole, 2006, Estes et al., 2007, Gathercole and Baddeley, 1989, Gathercole and Baddeley, 1990a, Gathercole and Baddeley, 1990b) and in children with SLI (Bishop et al., 1996, Briscoe et al., 2001, Leclercq et al., 2013, Munson et al., 2005). The articulatory complexity of consonant clusters stems from the fact that they require a range of complex and rapidly changing movements, rather than the simple oscillatory pattern of closing and opening the jaw entailed in producing a simple Consonant-Vowel (CV) sequence (MacNeilage and Davis, 2000, MacNeilage and Davis, 2005). Alternatively, the observed difficulty in repeating nonwords containing consonant clusters may simply result from their relative scarcity in the native languages of children tested in these studies (English, French), meaning that the children have had relatively less practice producing such sequences. Polish, the language of the current study, is very rich in consonant clusters (Dobrogowska, 1992) and thus provides a perfect ground for testing if repetition of nonwords containing consonant clusters remains difficult even when consonant clusters are frequent in the participant’s native language.
Another measure of phonological complexity that might reflect articulatory difficulty and thereby affect NWR is adherence to sonority-sequencing rules. Sonority is a feature of speech sounds, usually defined as the degree of opening of the mouth during the production of the sound (Jespersen, 1904, Keating, 1988) or the amplitude of the speech sound relative to other sounds produced in the same conditions (Parker, 2008, Selkirk et al., 1984, Steriade, 1982). Sonority-sequencing theories assume that a syllable is less phonologically complex if it begins with a less sonorous sound (e.g. an obstruent) and sonority rises steadily with each phoneme until it reaches its peak at a vowel (Dziubalska-Kołaczyk, 2002, Selkirk et al., 1984). For example, an initial consonant cluster /sp/, which consists of two obstruents (both low in sonority) violates sonority sequencing rules and thus is more complex than an initial cluster /pl/, which consists of an obstruent (low sonority) and a sonorant (high sonority). Sequences conforming to sonority rules tend to be easier to articulate, since such sequences typically require fewer articulatory gestures. The least articulatorily complex sequences are ones that conform to a simple pattern of closing and opening the jaw with minimal additional movement (MacNeilage and Davis, 2000, MacNeilage and Davis, 2005). Measures of articulatory complexity based on the sonority hierarchy (e.g. the number of sonority violations) have not, to the best of our knowledge, been used in research on NWR. The only exception is a study by Nimmo and Roodenrys (2002) who carried out a post-hoc analysis to determine whether phoneme sequences that conformed to the sonority hierarchy were more likely to be retained. However, the results of this study were inconclusive. Therefore, explicitly manipulating the sonority measure in an NWR study might provide information about the involvement of articulatory skills in the repetition of nonwords.
Phonologically complex nonwords may not only be more difficult to articulate but also more challenging to perceive. Simple consonant + vowel sequences tend to be more perceptually salient because vowels and consonants are acoustically distinct, which facilitates perception (Ohala & Kawasaki-Fukumori, 1997). Furthermore, vowels strengthen the acoustic cues to consonants, while two consonants occurring in sequence can mask each other’s acoustic cues, especially if the sequence violates sonority-based rules and contains two obstruents (Wright, 2004, Henke et al., 2012). So, if NWR performance depends on the phonological complexity of items (indexed by number of consonants or adherence to the sonority hierarchy), this could be because the task taps articulatory dexterity and/or perceptual acuity.
Clearly, repetition of a nonword entails perception of relevant details in input, temporary storage, and articulatory planning and production. Summarizing research to date, NWR tests have been proposed to tap into many different cognitive representations and mechanisms (phonological short-term memory, lexical representations, sublexical representations as well articulatory dexterity, with little mention of perceptual acuity), and empirical evidence has been advanced in support of these. However, the number and diversity of proposed factors and the cognitive processes associated with these poses problems for identifying sources of NWR performance. Although previous studies have pointed to the importance of these factors in NWR performance, they have largely ignored the fact that all the item parameters influencing NWR performance are likely to be correlated with one another, and therefore in part explain common variance. For example, lexical neighbourhood correlates with phonotactic probability (measured by phonemic bigram frequency), which leaves open the question which (one or both) truly helps nonword repetition (Messer et al., 2015, Metsala and Chisholm, 2010, Storkel et al., 2006, Vitevitch and Luce, 1998, Vitevitch and Luce, 1999, Vitevitch and Luce, 2005). Similarly, the presence of consonant clusters is correlated with phonotactic probability (consonant clusters are rare in English) and with subjective ratings of wordlikeness, at least in English-like nonwords (Coady and Evans, 2008, Gathercole et al., 1991). Likewise, length is highly correlated with neighbourhood density (the shorter the word or nonword, the more neighbours it has, as exemplified in Table 2). When sequences of nonwords are repeated, the effect of neighbourhood eclipses that of length (Jalberta, Neath, Bireta, & Surprenant, 2011) and it remains an open issue whether the same holds for nonword repetition.
The intercorrelation of the proposed factors highlights the first problem with previous approaches to investigating the processes of nonword repetition. Given the multitude of correlated factors, it is necessary to consider many item parameters in parallel in order to identify those that most directly explain the difficulty of repeating the nonword. However, previous studies have investigated only a few, and in most cases just one selected variable, precluding identification of the most direct predictors of NWR. Moreover, these studies often lacked sufficient statistical power to explore a larger number of item-related variables, due to the limited number of items (typically fewer than 30). Dichotomization of the selected (continuous) variable(s) further limits statistical power. In addition, the items used have often been highly specific, with length often limited to one syllable, and/or limited to one syllable structure (Gathercole et al., 1999, Messer et al., 2010, Messer et al., 2015, Roodenrys and Hinton, 2002, Thorn and Frankish, 2005, Vitevitch and Luce, 1998, Vitevitch and Luce, 1999, Zamuner et al., 2004, Zamuner, 2009).
A second shortcoming of previous studies is that some have evaluated a theoretical claim using indices that do not provide an adequate test of that claim. For example, support for theories positing that nonword repetition relies on sublexical representations hinged on the fact that phonemic bigram frequency predicts NWR accuracy. However, phonemic bigram frequency is a limited index of sublexical representation, providing no evidence regarding sublexical chunks greater than two phonemes. This limits evidence of the role of sublexical knowledge and implications for what is involved in NWR.
To go beyond previous studies and throw more light on key processes involved in NWR, we took a radically different approach to investigating factors influencing NWR performance. Making no a priori theoretical commitments, we set out to test the contribution of a wide array of item parameters using a much larger pool of items than those typically utilised. We targeted most of the predictors previously investigated, as well as including several new predictors that are theoretically justified (see Methods for details). The items were crafted in such a way that they spanned the entire space of possible values of the targeted item parameters.
In addition, our study targeted Polish rather than English. As Vitevitch, Chan, and Goldstein (2014) point out, testing English-based theories in languages other than English is crucial for advancing theories and evidence, and research on typologically different languages may be particularly informative. Polish provides an excellent source of evidence regarding the relative contribution of parameters such as syllable complexity (presence of consonant clusters) and phonotactic probability, because in contrast to English, consonants clusters are very frequent in Polish words. For example, English allows for 46 double consonant clusters and 11 triple consonants clusters word-initially (Trnka, 1966), while Polish allows for as many as 160 initial double clusters, around 100 initial triple clusters, as well as 20 quadruple clusters (Dobrogowska, 1992). This relates to the fact that Polish is a heavily consonantal language with 31 consonants, but only 6 oral vowels and 2 nasal diphthongs (in contrast, English has 24 consonants, but 12 vowels and 8 oral diphthongs; Roach, 2004, Gussmann, 2007, Jassem, 2003). Moreover, using Polish made it particularly important to check if the effect of sublexical knowledge on NWR is moderated by the knowledge of morphology, since Polish is a morphologically rich, inflectional language and Polish nonwords are likely to contain morphemes. For instance, Polish nouns are inflectionally marked for number and case (there are seven cases in Polish) and their declension depends on gender, with six gender classes in the language (Nagórko, 2007). There is also a rich declension system of adjectives and complex system of verb conjugation, with eleven conjugation classes, that produce a wide range of inflectional morphemes.
Our approach will help uncover the most important factors influencing nonword repetition, and thereby elucidate factors shaping the processing and learning of novel words. It will also have implications for the nature of deficits in children with SLI whose performance on NWR tasks is typically compromised.
Section snippets
Participants
Seventy-five children were recruited from four kindergartens in Kraków, Poland. All were monolingual speakers of Polish. Parents of all children signed informed consent and filled in a short questionnaire about parental education level, history of children's language and hearing problems, and ear infections. Ten children were excluded because they did not complete all tests (see below); six children because parents reported hearing deficits or serious ear infections; one child because he was
Results
On average, children correctly repeated 63.7% of the nonwords (range: 16–85%, SD: 15.0%). On average, items were correctly repeated by 64.5% of the children (range: 6.7–97.3%, SD: 20.4%). This shows that the demands of the items were appropriate to the children’s level of phonological development, with no floor or ceiling effects. At the same time, items varied widely in their overall difficulty, affording variance necessary for estimating predictors of repetition accuracy.
The best model
Discussion
In this study, we explored the mechanisms and representations underlying performance in the NWR task by testing the relationship between the scores on this task and potentially relevant item- and participant-related factors. Many of these factors have been considered in previous research and literature, but usually only individually. In this study, we explored them simultaneously, and making no a priori assumptions, in a large set of nonwords crafted in such a way as to span parameter space as
Conclusions
The current study is, to the best of our knowledge, the most comprehensive investigation of item-related predictors of NWR to date, and makes a number of novel contributions to the field. First, we propose a new index of sublexical support a nonword receives – average phonemic ngram frequency – which emerged as a much better indicator of nonword difficulty than other existing indices of phonotactic probability and lexical familiarity. We propose that this index should be used to assess
Acknowledgements
This work was initiated as a part of the COST Action ISO804. We thank Magdalena Smoczyńska for consulting material preparation, Ewa Haman for sharing details of the vocabulary test, and Marta Białecka-Pikul for providing support in accessing pre-schools and kindergartens to recruit participants. The construction of the nonword repetition test as well as data collection were supported by a subsidy from the Foundation for Polish Science to Z.W. During the work on the manuscript, J.S. and Z.W.
References (129)
- et al.
Determinants of wordlikeness: Phonotactics or lexical neighbourhoods?
Journal of Memory and Language
(2001) On the association between phonological memory and receptive vocabulary in five-year-olds
Journal of Experimental Child Psychology
(1996)- et al.
Modeling item length effects in memory span: No rehearsal needed?
Journal of Memory and Language
(1995) - et al.
Young children's sensitivity to probabilistic phonotactics in the developing lexicon
Journal of Experimental Child Psychology
(2004) - et al.
Evaluation of the role of phonological STM in the development of vocabulary in children: A longitudinal study
Journal of Memory and Language
(1989) - et al.
Phonological memory deficits in language disordered children: Is there a causal connection?
Journal of Memory and Language
(1990) - et al.
Memory for familiar and unfamiliar words: Evidence for a long-term memory contribution to short-term memory span
Journal of Memory and Language
(1991) - et al.
The relationships among verbal short-term memory, phonological awareness, and new word learning: Evidence from typical development and down syndrome
Journal of Experimental Child Psychology
(2009) - et al.
Specific language impairment: A deficit in grammar or processing?
Trends in Cognitive Sciences
(1998) - et al.
Phonology and syntax in specific language impairment: Evidence from a connectionist model
Brain and Language
(2003)